synth-ai 0.2.4.dev7__py3-none-any.whl → 0.2.4.dev9__py3-none-any.whl

synth_ai/environments/examples/crafter_classic/agent_demos/crafter_openai_ft/compare_experiments.py

@@ -0,0 +1,324 @@
+#!/usr/bin/env python3
+"""
+Compare experiments between gpt-4.1-nano and gpt-4.1-mini.
+Analyzes performance differences, achievement patterns, and instance difficulty.
+"""
+
+import json
+from typing import Any, Dict, List
+
+import duckdb
+import pandas as pd
+
+# Experiment IDs from the runs
+EXPERIMENTS = {
+    "gpt-4o-mini": "137683ed-3bd5-4bd3-9162-dae0371ddd3d",
+    "gpt-4o": "207307d5-4105-4a18-bb93-89936047fa18"
+}
+
+def connect_to_db():
+    """Connect to the DuckDB database."""
+    return duckdb.connect("synth_ai/traces/crafter_traces.duckdb")
+
+def get_experiment_summary(conn, experiment_id: str) -> Dict[str, Any]:
+    """Get basic experiment information."""
+    query = """
+    SELECT 
+        e.id,
+        e.name,
+        e.description,
+        e.created_at,
+        sv.branch,
+        sv.commit
+    FROM experiments e
+    LEFT JOIN experimental_systems es ON e.id = es.experiment_id
+    LEFT JOIN system_versions sv ON es.system_version_id = sv.id
+    WHERE e.id = ?
+    """
+    
+    result = conn.execute(query, [experiment_id]).fetchone()
+    if result:
+        return {
+            "experiment_id": result[0],
+            "name": result[1],
+            "description": result[2],
+            "created_at": result[3],
+            "branch": result[4],
+            "commit": result[5]
+        }
+    return None
+
+def get_session_stats(conn, experiment_id: str) -> pd.DataFrame:
+    """Get session-level statistics for an experiment."""
+    query = """
+    SELECT 
+        st.session_id,
+        st.created_at,
+        st.num_timesteps,
+        st.num_events,
+        st.num_messages,
+        st.metadata
+    FROM session_traces st
+    WHERE st.experiment_id = ?
+    ORDER BY st.created_at
+    """
+    
+    return conn.execute(query, [experiment_id]).df()
+
+def get_achievement_analysis(conn, experiment_id: str) -> Dict[str, Any]:
+    """Analyze achievements for an experiment."""
+    # Get session traces with achievement data
+    query = """
+    SELECT 
+        st.session_id,
+        st.metadata
+    FROM session_traces st
+    WHERE st.experiment_id = ? 
+    AND st.metadata IS NOT NULL
+    """
+    
+    results = conn.execute(query, [experiment_id]).fetchall()
+    
+    all_achievements = []
+    session_achievements = []
+    
+    for session_id, metadata in results:
+        if metadata:
+            # Parse the JSON metadata
+            try:
+                import json
+                metadata_list = json.loads(metadata) if isinstance(metadata, str) else metadata
+                
+                for meta_item in metadata_list:
+                    if isinstance(meta_item, dict) and meta_item.get('metadata_type') == 'SessionMetadum':
+                        data = meta_item.get('data', {})
+                        if 'achievements' in data:
+                            achievements_dict = data['achievements']
+                            num_achievements = data.get('num_achievements', 0)
+                            
+                            # Extract unlocked achievements
+                            unlocked = [ach for ach, unlocked in achievements_dict.items() if unlocked]
+                            all_achievements.extend(unlocked)
+                            
+                            session_achievements.append({
+                                'session_id': session_id,
+                                'num_achievements': num_achievements,
+                                'unlocked_achievements': unlocked,
+                                'total_achievements': len(achievements_dict)
+                            })
+            except Exception as e:
+                print(f"Error parsing metadata for session {session_id}: {e}")
+    
+    # Count achievements
+    achievement_counts = {}
+    for ach in all_achievements:
+        achievement_counts[ach] = achievement_counts.get(ach, 0) + 1
+    
+    return {
+        "total_achievements": len(all_achievements),
+        "unique_achievements": len(set(all_achievements)),
+        "achievement_counts": achievement_counts,
+        "achievement_list": all_achievements,
+        "session_achievements": session_achievements
+    }
+
+def get_model_usage_analysis(conn, experiment_id: str) -> pd.DataFrame:
+    """Analyze model usage and costs."""
+    query = """
+    SELECT 
+        e.model_name,
+        e.provider,
+        COUNT(*) as call_count,
+        SUM(e.prompt_tokens) as total_prompt_tokens,
+        SUM(e.completion_tokens) as total_completion_tokens,
+        SUM(e.total_tokens) as total_tokens,
+        SUM(e.cost) as total_cost,
+        AVG(e.latency_ms) as avg_latency_ms,
+        AVG(e.prompt_tokens) as avg_prompt_tokens,
+        AVG(e.completion_tokens) as avg_completion_tokens
+    FROM session_traces st
+    JOIN events e ON st.session_id = e.session_id
+    WHERE st.experiment_id = ? 
+    AND e.event_type = 'lm_cais'
+    GROUP BY e.model_name, e.provider
+    """
+    
+    return conn.execute(query, [experiment_id]).df()
+
+def get_session_performance_comparison(conn) -> pd.DataFrame:
+    """Compare session performance between experiments."""
+    query = """
+    SELECT 
+        st.experiment_id,
+        e.name as experiment_name,
+        COUNT(st.session_id) as total_sessions,
+        AVG(st.num_timesteps) as avg_timesteps,
+        AVG(st.num_events) as avg_events,
+        AVG(st.num_messages) as avg_messages,
+        SUM(st.num_timesteps) as total_timesteps,
+        SUM(st.num_events) as total_events,
+        SUM(st.num_messages) as total_messages
+    FROM session_traces st
+    JOIN experiments e ON st.experiment_id = e.id
+    WHERE st.experiment_id IN (?, ?)
+    GROUP BY st.experiment_id, e.name
+    ORDER BY e.name
+    """
+    
+    return conn.execute(query, [EXPERIMENTS["gpt-4o-mini"], EXPERIMENTS["gpt-4o"]]).df()
+
+def get_achievement_comparison(conn) -> pd.DataFrame:
+    """Compare achievements between experiments."""
+    # This is a more complex query to extract achievements from metadata
+    query = """
+    WITH achievement_data AS (
+        SELECT 
+            st.experiment_id,
+            e.name as experiment_name,
+            st.session_id,
+            e.metadata,
+            e.event_metadata
+        FROM session_traces st
+        JOIN experiments e ON st.experiment_id = e.experiment_id
+        JOIN events ev ON st.session_id = ev.session_id
+        WHERE st.experiment_id IN (?, ?)
+        AND ev.event_type = 'environment'
+        AND ev.metadata IS NOT NULL
+    )
+    SELECT 
+        experiment_id,
+        experiment_name,
+        COUNT(DISTINCT session_id) as sessions_with_achievements,
+        COUNT(*) as total_achievement_events
+    FROM achievement_data
+    GROUP BY experiment_id, experiment_name
+    """
+    
+    return conn.execute(query, [EXPERIMENTS["gpt-4o-mini"], EXPERIMENTS["gpt-4o"]]).df()
+
+def analyze_instance_difficulty(conn) -> Dict[str, Any]:
+    """Analyze which instances were more difficult for each model."""
+    query = """
+    SELECT 
+        st.experiment_id,
+        e.name as experiment_name,
+        st.session_id,
+        st.num_timesteps,
+        st.num_events,
+        st.metadata
+    FROM session_traces st
+    JOIN experiments e ON st.experiment_id = e.id
+    WHERE st.experiment_id IN (?, ?)
+    ORDER BY st.experiment_id, st.session_id
+    """
+    
+    df = conn.execute(query, [EXPERIMENTS["gpt-4o-mini"], EXPERIMENTS["gpt-4o"]]).df()
+    
+    # Group by experiment and analyze session patterns
+    analysis = {}
+    for experiment_id in [EXPERIMENTS["gpt-4o-mini"], EXPERIMENTS["gpt-4o"]]:
+        exp_data = df[df['experiment_id'] == experiment_id]
+        analysis[experiment_id] = {
+            "total_sessions": len(exp_data),
+            "avg_timesteps": exp_data['num_timesteps'].mean(),
+            "avg_events": exp_data['num_events'].mean(),
+            "max_timesteps": exp_data['num_timesteps'].max(),
+            "min_timesteps": exp_data['num_timesteps'].min(),
+            "session_lengths": exp_data['num_timesteps'].tolist()
+        }
+    
+    return analysis
+
+def main():
+    """Main analysis function."""
+    print("🔍 COMPARING GPT-4O-MINI vs GPT-4O EXPERIMENTS")
+    print("=" * 80)
+    
+    conn = connect_to_db()
+    
+    # Get experiment summaries
+    print("\n📋 EXPERIMENT SUMMARIES")
+    print("-" * 40)
+    
+    for model_name, exp_id in EXPERIMENTS.items():
+        summary = get_experiment_summary(conn, exp_id)
+        if summary:
+            print(f"\n{model_name.upper()}:")
+            print(f"  Name: {summary['name']}")
+            print(f"  ID: {summary['experiment_id']}")
+            print(f"  Created: {summary['created_at']}")
+            print(f"  Git: {summary['branch']} @ {summary['commit'][:8]}")
+    
+    # Session performance comparison
+    print("\n📊 SESSION PERFORMANCE COMPARISON")
+    print("-" * 40)
+    
+    perf_df = get_session_performance_comparison(conn)
+    print(perf_df.to_string(index=False))
+    
+    # Achievement analysis
+    print("\n🏆 ACHIEVEMENT ANALYSIS")
+    print("-" * 40)
+    
+    for model_name, exp_id in EXPERIMENTS.items():
+        print(f"\n{model_name.upper()}:")
+        achievement_data = get_achievement_analysis(conn, exp_id)
+        print(f"  Total Achievements: {achievement_data['total_achievements']}")
+        print(f"  Unique Achievements: {achievement_data['unique_achievements']}")
+        print(f"  Achievement Counts: {achievement_data['achievement_counts']}")
+    
+    # Model usage analysis
+    print("\n💰 MODEL USAGE ANALYSIS")
+    print("-" * 40)
+    
+    for model_name, exp_id in EXPERIMENTS.items():
+        print(f"\n{model_name.upper()}:")
+        usage_df = get_model_usage_analysis(conn, exp_id)
+        if not usage_df.empty:
+            print(usage_df.to_string(index=False))
+        else:
+            print("  No model usage data found")
+    
+    # Instance difficulty analysis
+    print("\n🎯 INSTANCE DIFFICULTY ANALYSIS")
+    print("-" * 40)
+    
+    difficulty_analysis = analyze_instance_difficulty(conn)
+    
+    for model_name, exp_id in EXPERIMENTS.items():
+        data = difficulty_analysis[exp_id]
+        print(f"\n{model_name.upper()}:")
+        print(f"  Total Sessions: {data['total_sessions']}")
+        print(f"  Avg Timesteps: {data['avg_timesteps']:.1f}")
+        print(f"  Avg Events: {data['avg_events']:.1f}")
+        print(f"  Timestep Range: {data['min_timesteps']} - {data['max_timesteps']}")
+    
+    # Performance comparison summary
+    print("\n📈 PERFORMANCE COMPARISON SUMMARY")
+    print("-" * 40)
+    
+    mini_data = difficulty_analysis[EXPERIMENTS["gpt-4o-mini"]]
+    full_data = difficulty_analysis[EXPERIMENTS["gpt-4o"]]
+    
+    print(f"GPT-4O-MINI:")
+    print(f"  Sessions: {mini_data['total_sessions']}")
+    print(f"  Avg Timesteps: {mini_data['avg_timesteps']:.1f}")
+    print(f"  Avg Events: {mini_data['avg_events']:.1f}")
+    
+    print(f"\nGPT-4O:")
+    print(f"  Sessions: {full_data['total_sessions']}")
+    print(f"  Avg Timesteps: {full_data['avg_timesteps']:.1f}")
+    print(f"  Avg Events: {full_data['avg_events']:.1f}")
+    
+    # Calculate improvements
+    timestep_improvement = ((full_data['avg_timesteps'] - mini_data['avg_timesteps']) / mini_data['avg_timesteps']) * 100
+    event_improvement = ((full_data['avg_events'] - mini_data['avg_events']) / mini_data['avg_events']) * 100
+    
+    print(f"\n📊 IMPROVEMENTS:")
+    print(f"  Timesteps: {timestep_improvement:+.1f}%")
+    print(f"  Events: {event_improvement:+.1f}%")
+    
+    conn.close()
+
+if __name__ == "__main__":
+    main()